Document resource
Introduction Capsule endoscopy (CE) provides a non-invasive approach to visualise the small bowel mucosa and assess inflammatory activity in Crohn’s disease (CD), AI-assisted software within CE is increasingly being used to support reporting. 1 This study compares AI derived analysis of CE videos using the OMOM software with expert reads to determine whether AI scores accurately reflect disease severity and inform clinical management.Methods A single centre cohort study was conducted including patients with suspected or established CD identified from the OMOM CE database. Clinical data were retrospectively collected, including faecal calprotectin, laboratory parameters and disease activity assessed using the Harvey Bradshaw Index. All CE videos underwent standard read (SR) by one of two independent expert readers, including assessment of bowel preparation. The same reader, blinded to the previous read, re-analysed the studies using AI assisted CE software to produce a hybrid read (HR). Lewis scores and reading times were calculated for all CE videos for both SR and HR. Agreement between readings were evaluated using a mixed effects interclass correlation co-efficient. Prospective data on subsequent management decisions and further investigations were recorded to explore the clinical impact of CE findings on patient care.Results The study included 24 patients, with 2 excluded due to incomplete CE; 14 (58.3%) underwent CE for known CD and 10 (41.7%) for suspected CD. Median age was 38.5 years (IQR 30.5-51.3) and mean small bowel transit time was 209m29s (± 70m32s). Among patients with available clinical data, anaemia was present in none, C-reactive protein was elevated in 4 (16.7%), and the median FC was 85.5µg/g (IQR 21.0-235.3). Agreement between expert SR and HR for overall Lewis scores was good (ICC 0.925, 95% CI 0.828–0.968; p<0.001), with perfect agreement in tertile 1 and good to moderate agreement in tertiles 2 and 3 (ICC 0.886 and 0.694, respectively; p<0.001). Follow-up demonstrated management changes in 5 patients (20.8%) with Lewis scores >500 on both HR and SR. One patient (4.2%) underwent double-balloon enteroscopy following CE, with management decisions still under review in a few cases. The mean reading time for SR was 25m and 6m56s for HR.Conclusion AI assisted CE analysis using the OMOM software has demonstrated good agreement with expert human assessment of disease activity in CD, and identified patients requiring changes in management. However, variability in agreement across tertiles indicates that expert oversight remains essential and AI assisted interpretation should be used as an adjunct rather than a replacement for standard CE reading.Reference Ding Z, Shi H, Zhang H, et al. Gastroenterologist-level identification of small-bowel diseases and normal variants by capsule endoscopy using a deep-learning model. Gastroenterology 2019;157:1044–1054 e5.